Early and visible detection of fugitive methane leaks in industrial infrared videos via local-global contrast enhancement

Early visualization of greenhouse gases and volatile organic compounds is essential for emission control and industrial safety. Infrared optical gas imaging sequences are difficult to interpret because plumes occupy little of each frame. In our data, mean plume coverage was 6.12%, mean local contrast was 8.09%, and over 72% of samples had local contrast below 10%. We develop an artificial intelligence framework for infrared methane-plume enhancement and early leak detection. It embeds a Local Perception Enhancement Module in a U-shaped Transformer and combines differentiable local contrast limiting, spatial gating, and residual fusion. The study used 180,000 infrared frames from controlled releases and an operating chemical industrial park. The framework achieved a region-of-interest peak signal-to-noise ratio of 35.41 dB and a global structural similarity index of 0.951. Structural ablation confirmed gains in local and global fidelity, structural similarity, and contrast. With a fixed one-stage object detector, enhanced inputs increased mean intersection over union from 76% to 85% and reduced false negatives by 31.5%. For controlled releases of 80 to 4000 mL per minute, enhancement lowered the flow rate at which probability of detection reached 0.9 from 1000 to 300 mL per minute. At two field sites, stable detection occurred 10.5 and 11.8 s earlier. These results connect artificial intelligence-based image enhancement to measurable gains in industrial methane-leak monitoring.